This is classic way of reducing CPU usage, just bounce a part of a track to raw audio an play it back so it doesn't need to render in real time. A GPU doesn't really change the equation there.
There are some methods of synthesis which rely on FFT which can't really be done well in real-time with the CPU (PadSynth, PaulStretch) that I'm hoping this will help with.